Intelligent anatomy of spinal tuberculosis: region-specific performance analysis of a deep learning model based on multi-center CT axial images

Objective To solve the problems of insufficient model generalization ability and unclear influence of anatomical regional differences in the intelligent diagnosis of spinal tuberculosis, and to construct a multi-center CT deep learning model.Method This study adopted a retrospective multi-center design and collected CT images of 1025 patients with spinal tuberculosis from 4 hospitals in the southwest region from 2012 to 2024. In the preprocessing stage, DICOM is converted to 224x224 RGB format and marked by senior radiologist. A binary classification model was constructed based on ResNet50 transfer learning and trained for 100 rounds using the Adam optimizer (lr = 1e-4, batch = 8). The verification framework is divided into three-stage validation framework: Center 1, 80% training of data/verification within 20%; Centers 2 to 4 were combined into the external test set and subdivided by cervical vertebrae/thoracic vertebrae/lumbar vertebrae/sacrum vertebrae. The evaluation indicators include AUC, accuracy, precision, recall and F1 score, supplemented by Grad-CAM visualization.Result The internal validation performance of the model was excellent (AUC = 0.90), but the external test decreased (AUC = 0.88). Anatomical stratification showed that the cervical vertebrae were the weakest (AUC = 0.70), the lumbar vertebrae were the most stable (AUC = 0.91), the sacral vertebrae had high sensitivity (recall rate = 0.756), and the thoracic vertebrae had a high false positive rate.Conclusion The anatomical dependence of diagnostic performance was revealed for the first time, and tiered deployment strategy is suggested: priority deployment of high-stability segments (lumbar vertebrae/sacrum), low-efficiency regions (cervical vertebrae) combined with MRI verification, and hierarchical data standards were established to optimize the algorithm.